'and' joining parallel objects
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 15 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron lights up on the word 'and' or 'or' when it links two things that share the same role, such as 'understanding of and applying' or 'control of and manage'. When it was turned up, the model's text became repetitive and drifted off topic, so the steering showed no clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
Mudejar at practically all levels of the economy, and the quotidian mingling of Muslim and Christian in field, workshop, and marketplace provided the foundation of—and the most effective adhesive reinforcing—the structure of Valencia's plural society. A common pre
-
time should focus on four critical areas: (1) developing understanding of and applying proportional relationships; (2) developing understanding of operations with rational numbers and working
-
/conflicts are most familiar to them.⏎Do the standards support students’ engagement in ongoing self-examination to excavate how one’s identities inform their understandings of and experiences with complex social problems (Mitchell, 20
-
the average volume of Internet traffic has historically grown by 40–50% per year. This growth leads to the growth of the transit fees paid by and paid to the upstream ISPs.⏎For example, while the price for Internet Transit drops
-
to protect millions of acres of forest and keep millions of tons of CO2 in the ground.⏎Our grants support organizations and communities that are working to regain control of and sustainably manage their traditional territories through land title initiatives, community education, development of
-
did not get any education on that part of their country and it hardly ever surfaces as part of the political discussion. Despite the active participation of Nubians in and before the revolution; their efforts to highlight their cause and their history of discrimination,
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: if you don't think about the other side in the "other" side of the mind and you are thinking only about your own, you are not thinking. This is the theory of the "mind of the self and the mind of the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
⏎- A pair of matching socks⏎- Something that is soft⏎- Something that is round⏎- A photo of someone you love⏎- Something you love to play with<|endoftext|>The report’s focus is on perceptions of changes among pastoralists and
-
photograph of our fascinating universe is featured, along with a brief explanation written by a professional astronomer.⏎2013 May 19⏎Explanation: What on Earth is that? The Richat Structure in the Sahara Desert of Mauritania is
-
, a person’s views, or their behavior, all of which are not within your control? Is there a power struggle in the making? What exactly is triggering to you? When you’re aware of it, you can decide how to handle it
-
the adherent to a religion is necessarily, from the scholastic point of view, the proper person to speak for the religion. They go on to characterize as a "tightrope between academic and religious worlds of value" the situation of
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L14:2627 excites 'or' or 'and' joining two items
- L3:1114 suppresses
- L7:669 weakly excites
- L13:2255 weakly suppresses
- L3:537 weakly suppresses
- L14:9 weakly suppresses
- L1:1624 weakly suppresses
- L7:1784 weakly excites punctuation right after 'of' or 'with'
- L10:2522 weakly excites
- L7:68 weakly suppresses
- L11:190 weakly excites
- L11:2877 weakly suppresses
- L14:5 weakly suppresses
- L6:1018 weakly suppresses
- L9:2244 weakly excites
- L11:519 weakly suppresses
- L13:2069 weakly suppresses
- L12:340 weakly suppresses
- L14:868 weakly suppresses
- L14:2249 weakly excites
- L2:36 weakly excites
- L11:2288 weakly excites
- L11:807 weakly suppresses
- L4:628 weakly suppresses
Feeds later layers
- L18:2703 excites
- L19:2327 excites
- L17:1679 excites
- L20:2454 weakly excites
- L17:2067 weakly excites
- L17:2017 weakly excites
- L25:477 weakly excites
- L18:2557 weakly excites
- L16:2537 weakly excites
- L18:2255 weakly excites
- L16:2912 weakly excites
- L21:1757 weakly excites
- L18:834 weakly excites
- L17:356 weakly excites
- L17:294 weakly excites
- L19:504 weakly excites
- L17:1886 weakly excites
- L24:55 weakly excites
- L23:1317 weakly excites
- L16:1192 weakly excites
- L22:793 weakly excites
- L21:621 weakly excites
- L16:247 weakly excites
- L25:1132 weakly excites
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- 嗑
- 必备
- ilyn
- ursor
- inity
- ina
- iti
- 许
- ят
- /Input
Suppresses
- Wise
- いない
- xp
- 骢
- 端
- CastException
- anson
- urgeon
- 鼓
- XP